https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106470
Bug ID: 106470
Summary: Subscribed access to __m256i casted to (uint16_t *)
produces garbage or a warning
Product: gcc
Version: 12.1.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: vt at altlinux dot org
Target Milestone: ---
It's appeared in liboqs in code imported from PQClean[1]. There is minimized
reproducer:
1.
```
$ cat test1.c
#include <immintrin.h>
#include <stdio.h>
#include <stdint.h>
#include <string.h>
int main(void)
{
__m256i tmp = _mm256_set_epi16(1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12,
13, 14, 15, 16);
for (size_t i = 0; i < 16; i++) {
printf(" %04x", ((uint16_t *)&tmp)[i]);
}
printf("\n");
return 0;
}
$ gcc -O2 -Wall -mavx2 -o a test1.c
$ ./a
c040 69a1 558e 0000 68da b7f6 7fb1 0000 87d0 b7f6 7fb1 0000 0000 0000 0001
0000
```
This should been printing:
0010 000f 000e 000d 000c 000b 000a 0009 0008 0007 0006 0005 0004 0003 0002
0001
2. If we add `#pragma GCC unroll 16` it compiles with the warning, still
printing incorrect values.
```
$ cat test2.c
#include <immintrin.h>
#include <stdio.h>
#include <stdint.h>
#include <string.h>
int main(void)
{
__m256i tmp = _mm256_set_epi16(1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12,
13, 14, 15, 16);
#pragma GCC unroll 16
for (size_t i = 0; i < 16; i++) {
printf(" %04x", ((uint16_t *)&tmp)[i]);
}
printf("\n");
return 0;
}
$ gcc -O2 -Wall -mavx2 -o b test2.c
test2.c: In function ‘main’:
test2.c:10:51: warning: ‘tmp’ is used uninitialized [-Wuninitialized]
10 | printf(" %04x", ((uint16_t *)&tmp)[i]);
| ~~~~~~~~~~~~~~~~~~^~~
test2.c:7:17: note: ‘tmp’ was declared here
7 | __m256i tmp = _mm256_set_epi16(1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
11, 12, 13, 14, 15, 16);
| ^~~
$ ./b
0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000
0000
```
In comparison, clang produces correct output.
[1] https://github.com/open-quantum-safe/liboqs/issues/1244
